The Pentagon would supply at most a quarter of the debt behind Fluidstack's data centers

The Office of Strategic Capital's first credit program financed equipment for domestic manufacturing ^[6]. Equipment is the easy case for a lender: the asset has a serial number and a used market. You can repossess a machine tool; a substation stays where it was built. An AI data center needs land, grid connections, substations, cooling equipment, networking systems and large quantities of accelerators, and developers frequently commit capital years before a site produces revenue ^[7].

Now the size. Forbes, reading UK filings, put Fluidstack's 2024 revenue above $66 million ^[9]. Five billion dollars is about 76 times that: 5,000 divided by 66 is 75.8 ^[1]. It is roughly 3.3 times the $1.5 billion round Jane Street led at a valuation above $18 billion, which followed a reported $750 million financing led by Situational Awareness at a $7.5 billion valuation ^[10][3]. The proposed loan would exceed either equity round ^[12].

The debt side is where the framing of federal credit as a primary funding source breaks. Forbes reported, from regulatory filings, that Fluidstack and its partners had accumulated more than $15 billion of debt for infrastructure projects, and Fluidstack has not confirmed all of those terms ^[11]. Add $5 billion and the federal slice is at most about a quarter of a roughly $20 billion total: 5 divided by 20 ^[2]. Private lenders have already committed more than three times the amount under discussion ^[4]. Congress authorized the office to lend and guarantee across 31 covered technology categories ^[5].

An investor memo cited by Forbes claimed Fluidstack could build a data center in three months, compared with at least a year for Google, Amazon or Meta, and projected the company would manage up to 1.3 gigawatts across more than 10 sites this year ^[14]. Divide the projection and the average site is at most about 130 megawatts ^[5]. For the three-month figure to mean anything to a lender, the clock would have to start where the lender's money starts, and the memo as reported does not say where it starts. The founders are trying to convert customer commitments and power agreements into operating data centers, having moved from aggregating compute into developing and operating capacity ^[19]. Wu and Cox began in 2017 at Oxford by pooling GPU gaming machines that sat unused most of the week for researchers who could not get compute ^[15]. Construction schedules, cooling and power delivery decide whether contracted hardware produces revenue ^[16].

Which projects qualify for the proceeds matters more than the headline number. A loan secured against contracted data-center revenue carries different risks from financing capacity built before customers commit ^[18]. Anthropic selected Fluidstack to build custom data centers as part of Anthropic's announced $50 billion US infrastructure investment ^[13], so at least part of the pipeline has a named offtaker. The interest rate, collateral, maturity and covenants have not been reported, and those terms decide whether federal credit is longer or cheaper than what the private debt markets would charge ^[8].